Use casesSupermarket
…Cameras —New rule

Ask your store anything

Ask what happened on the floor in plain words. Muse Spark answers from the event log and the video; SAM 3.1 confirms a shelf gap holds no product.

Shoppers unloading trolleys at a supermarket checkout lane
Stock footage, Pexels video 39221979 (faces blurred): the checkout clip the Ask agent watched on 1 Nov.

Supermarket

Your cameras already see the queue, the empty shelf and the spill. Ask anything in your own language; Muse Spark answers from what SAM 3.1 tracked and the Tier 1 model saw, turns the answer into a change you can test, watches for it, tells the right person with the photo, checks the fix twice and keeps a receipt. Everything below is live on this store’s cameras.

Connecting to the store’s cameras…

What one question proved

REPLAYrecorded 4 Nov 2026: A recorded run, replayed with no model calls

22Ask's answer before: the tracker's count

10people

The checkout queue camera, one frame at 01:13:07 UTC. Counted by hand: about 12.

Three counts of the same frame, none seeing the others:

“SAM found 15 boxes but 9 had no person point, indicating duplicates.”

SAM 3.115-17 boxesMuse Spark 1.39 pointsMuse Glimmer 30Bcounted 10-11

Ask's error went from +10 and +8 to -2 and -2 on two busy frames.

A spill, seen and cleared in 1.9 min

SAMPLEthe spill photo is a Muse Image edit: Hand-written sample content, not a run or a measured result
  1. Seen on the aisle camera

    Muse Spark found it, SAM 3.1 boxed the puddle, Glimmer agreed

  2. Staff alerted

    Chain record #19

  3. Cleared, 1.9 min after it was seen

    Proof photo checked by Spark (9.8 s) and Glimmer (15.1 s); chain record #21

Chain intact that day: 28 records.

The same frames, before and after

Frame (UTC)By handBeforeAsk now
01:13:07~122210
01:03:10~132111

No single model was right: SAM over-counts busy views and Spark's points miss people half hidden; the three counts bracket the truth.

Source: reports/eval/2026-11-04-supermarket-live.md

From a question to a fixed problem

  1. Step 1, not yet:

    Ask

    Typed or spoken, any language; the answer is read aloud.

  2. Step 2, not yet:

    Advice

    A card with a pre-registered check; its numbers must be in the answer's evidence.

  3. Step 3, not yet:

    Watch

    Queue, hazard and shelf checks tell the right person, with the photo.

  4. Step 4, not yet:

    Fixed

    A staff photo checked by Spark and the Tier 1 model closes the task; the time goes on the chain.

  5. Step 5, not yet:

    Brief

    Muse Spark writes the owner's brief from the stored checks and tasks.

1

Ask

typed or spoken, any language; answer read aloud
2

Advice with a pre-registered check

Site Advisor card from the answer; numbers must appear in the answer's evidence

Ask a question first (step 1).

3

The watch tells staff, with the photo

queue / hazard / shelf checks; alerts go to the staff alerts app (and Telegram or WhatsApp when set up)
CameraRecordedCheck now
Standing watches

None yet: accept an advice card or add one.

4

Fixed, and proven fixed

a staff photo checked by Spark and the Tier 1 model, or a clean camera re-check; time to fix goes on the evidence chain

No staff tasks yet: a shelf or hazard check that finds something opens one.

5

Owner’s brief

written by Muse Spark from the stored checks, tasks, questions and the running test
Chain scorecard row